v0.1.3 — Configurable autocompactor + polish
The chat now ships an end-to-end autocompact gate at the top of SuccessorChat._begin_agent_turn. The gate reads the active profile's new CompactionConfig, builds a ContextBudget against the resolved context window, and (when crossing the threshold) defers the agent turn behind a background compaction worker that resumes the turn against the freshly compacted log.
Highlights
- Configurable autocompactor. Percentage-based thresholds you can tune per profile, with hard token floors so tiny windows still get usable headroom. Setup wizard exposes four presets (default / aggressive / lazy / off); the config menu has a full per-field editor.
- Post-compact size assertion. If the new log is at least 90% of the original,
compact()stamps a warning on the boundary marker and the chat shows anunderperformedannotation. - 974 tests passing. 93 new tests covering the dataclass, percentage scaling at 8K → 2M windows, the chat-layer gate, the post-compact assertion, edge cases, and visual snapshot tests for the wizard step + config menu section.
- Default profile now ships with bash enabled. New users see the agentic loop on their first turn.
successor-devprofile uses the aggressive preset so harness-development sessions stay responsive at the edge of the context window.
Visuals
The release assets include 7 GIFs cropped from the walkthrough recordings, showing the renderer + the chat in action against a real model:
intro_braille.gif— the custom braille intro animationwizard_theme.gif— wizard theme cycling between steel and forge redbraille_red.gif/braille_blue.gif— the SUCCESSOR braille emergence in two themestool_dispatch.gif— multi-tool heredoc withfind-filesandgit-logrunning back to backsearch_demo.gif— conversation search with highlighted matcheschat_streaming.gif— model streaming with thinking indicator and aread-filetool dispatch
Docs
- New
docs/compaction.mdcovers the schema, threshold math, gate flow, and the failure modes the post-compact assertion catches - README has an Inspirations section crediting Cheng Lou's Pretext, Hermes Agent + the open-source agent harness ecosystem, and the broader open-source AI community
Full release notes in CHANGELOG.md.